Performance Co m pu t i ng Because of the mismatch between 110 and CPU speeds, highperformance computers have long been forced to confront the fundamental 110 bottleneck. As processing power and memory size continue to grow rapidly for micro-and minicomputers, they too will become 110 limited. A number of hardware and software approaches such as parallel read-out disks, expanded storage (for example, solid-state disks), and disk striping, have been used to increase //O bandwidth and thus narrow

more »

... the CPU-I/O performance gap. In addition, new developments driven by advances in smalldiameter (5.25-in and 3.5-in) disk drives, promise very high 110 bandwidth if large numbers of devices can be organized into arrays of disks. In this paper we review the state of the art in disk devices and 110 controllers, and will describe new approaches for very highperformance /IO based on redundant arrays of inexpensive disks (RAIDS). I. INTRODUCTION Architects of high-performance computers have long been forced to acknowledge the existence of a large gap between the speed of the CPU and the speed of its attached I/O devices. We review a number of techniques that have been developed in an attempt to narrow this gap. The computational environment for high-performance computing, however, is undergoing radical changes. The distinction between supercomputers, minisupercomputers, superm i n icom puters, and h igh-performance workstat ions is rapidly becoming blurred. For example, the recently announced Intel i860 processor chip possesses the same processing power as the original CRAY-1 (approximately 33 scalar MIPS), yet sells in OEM quantities at $750. The concept of the diskless supercomputer, that is a high-performance computational machine connected via high-speed network to a file server with large numbers of disk devices, is becoming a possibility. This presents new challenges for system architects sincethe VOsystem is likelyto beattached Manuscript